How We Built NASA's Zero-Shot Detector for Gas Leaks from Space

A deep learning pipeline was developed at NASA JPL to automatically detect methane and other gas plumes from space using data from the EMIT instrument. The system combines a matched filter for feature extraction, a U-Net model for semantic segmentation, and physics-informed scoring to ensure high accuracy and reduce false alarms. A key innovation is its zero-shot generalization, where the model, trained only on methane patterns, can identify other gases like carbon monoxide and ammonia without retraining. This automated system has drastically reduced detection latency and discovered numerous previously missed emission sources in historical data.